vfs: handle vfs_init() failures

[kevans]

Summary

Most vfs_init implementations will not fail, with the notable current
exception that tmpfs_subr_init() can fail to allocate a new swap pager
type, in which case we probably do not want to proceed and keep it
registered.

Another commit will also start to propagate a fatal error during
linsysfs init, which should likely also not keep the vfs registered.

[gbe]

Small typo in the comment.

sys/kern/vfs_init.c

523

s/ony/only/

[kib]

sys/kern/vfs_init.c

523

I do not understand the motivation for this. IMO we should unwind always. Memorize the state of the vfsconf before loading, and reset ot.

Also, I think there should be something that prevents parallel registration/unwind/unregistration.

534

[kevans]

sys/kern/vfs_init.c

523

Both vfs_register and vfs_unregister takes the vfsconf_sx exclusively, and for the former it's held up to here. That seemed sufficient, is there some scenario where it isn't?

[kib]

sys/kern/vfs_init.c

523

Sure, I jujst forget about them.

As an excuse, I can say that it was me who added vfsconf_lock (>10 years ago).

[kevans]

sys/kern/vfs_init.c

523

Ah, ok, perfect. =)

I'll fix your other concerns and update, since we might still want to do this for the off-chance that tmpfs fails to allocate a swap pager type (since linsysfs init will get pared down as you suggested).

[kevans]

Address review feedback: always unwind maxvfsconf change, compare != 0

[fuz]

Kernel now boots on the Windows ARM Dev Kit. LGTM.